US Food Nutrition Composition
Every food in USDA's authoritative composition databases with the amount of every nutrient it contains, in one clean long table - the join nobody wants to do by hand. FoodData Central ships this as five separate CSVs (food, food_nutrient, nutrient, food_category and a per-type NDB crosswalk); we assemble and de-duplicate them into one tidy (food × nutrient) Parquet. Covers the three pure-USDA bundles - Foundation Foods (analytically measured reference foods), SR Legacy (the classic ~7.8k-food Standard Reference that underpins US nutrition analysis) and FNDDS (the 'as consumed' survey foods people actually report eating) - with energy, macronutrients, vitamins, minerals, amino acids and fatty acids, each with its unit and amount per 100 g. A data_type column preserves the measured-vs-survey distinction so you can filter either way. Keyed by FDC id with the legacy NDB number carried through, so it joins to recipe, dietary-intake and agricultural data. Public domain (U.S. Government work).
- Source
- USDA FDC
- Licence
- Redistributable open data
- Quality
- QC clean · 0 failures
- Format
- Apache Parquet + dictionary
Safe to train on
Built only from a named, redistributable official source under a documented open licence - not scraped web data. No copyright grey zone, no personal data. Ships machine-readable Croissant metadata (ML Commons - loads in Hugging Face / Kaggle / Google), an AI training-licence manifest documenting source, licence and provenance for your model's data governance, and a machine-readable data dictionary (drop it into an agent / RAG prompt so the model knows every column) - all inspectable before you buy.
What's included
- One row per (food × nutrient) - the full nutrient panel for every food, tidy long shape
- Foundation + SR Legacy + FNDDS (15k+ foods): energy, macros, vitamins, minerals, amino & fatty acids
- data_type flags measured (foundation/sr_legacy) vs 'as consumed' survey (survey_fndds_food) foods
- Amount per 100 g with the nutrient's unit; nutrient rank for canonical ordering
- Keyed by FDC id; legacy NDB number carried through for joins to older nutrition datasets
- The multi-CSV join (food ⋈ food_nutrient ⋈ nutrient ⋈ category ⋈ NDB) done once, cleanly
- Branded foods deliberately excluded (manufacturer-label provenance) - pure USDA data only
- Public domain (USDA FoodData Central) - food-composition facts only, no personal data
- Source: U.S. Department of Agriculture, FoodData Central (public domain)
Schema
| Column | Type | Description | Filled | Distinct |
|---|---|---|---|---|
| fdc_id | int | FoodData Central food id (primary key with nutrient) | 100.0% | 13,611 |
| ndb_number | str | Legacy NDB / Standard-Reference number (where assigned) | 64.6% | 7,924 |
| description | str | Food name / description | 100.0% | 13,374 |
| food_category | str | USDA food category (or WWEIA survey category for FNDDS foods) | 100.0% | 196 |
| data_type | str | Source bundle: foundation_food, sr_legacy_food or survey_fndds_food | 100.0% | 3 |
| nutrient | str | Nutrient name (e.g. Protein, Energy, Calcium, Ca) | 100.0% | 245 |
| nutrient_unit | str | Unit of the amount (G, MG, UG, KCAL, …) | 100.0% | 7 |
| amount | float | Amount of the nutrient per 100 g of edible portion | 100.0% | 24,223 |
| nutrient_rank | float | USDA display/sort rank for the nutrient | 100.0% | 239 |
Sample & preview
Every purchase ships as Apache Parquet with a data dictionary and the full QC report. A free sample (first rows + schema) is downloadable here - confirm fit before you buy. Source: USDA FDC (redistributable open data; attribution passes through - see our Licence).
Don't trust screenshots - drop the sample into your notebook right now.
import pandas as pd df = pd.read_parquet("us-food-nutrition-composition_sample.parquet") df.info() # typed columns, gap-aware, ready to join⬇ Download free Parquet sample or CSV for Excel
Related datasets
European Health System
The health-system profile of every European country in one clean panel: life expectancy at birth, hospital beds and practising physicians per 100 000 inhabitants, and total health expenditure as a share of GDP. Cleaned from Eurostat health statistics into one tidy annual per-country table - the capacity-and-outcomes view health-policy, insurance and life-science teams benchmark against.
US Drug Adverse
How abnormal each drug's adverse-event reporting is now: the z-score of monthly FAERS counts against the drug's own trailing 24-month baseline. |z| ≥ 3 = a 3σ safety-signal surge. Pre-computed from the FAERS monthly-counts product.
US Drug Adverse
Monthly counts of FDA adverse-event reports (FAERS) for ~40 major drugs by generic name - statins, GLP-1s, anticoagulants, biologics, oncology and more. Built from openFDA's COUNT endpoint, so it carries zero individual reports and zero patient data: a clean safety-signal time series per drug.